feat(BA-2213): Implement API Layer of Model Deployment

changes/5672.feature.md

@@ -0,0 +1 @@
+Implement API Layer of Model Deployment

docs/manager/graphql-reference/supergraph.graphql

@@ -3009,10 +3009,10 @@ type ModelDeployment implements Node
   metadata: ModelDeploymentMetadata!
   networkAccess: ModelDeploymentNetworkAccess!
   revision: ModelRevision
-  scalingRule: ScalingRule!
-  replicaState: ReplicaState!
   defaultDeploymentStrategy: DeploymentStrategy!
   createdUser: UserNode!
+  scalingRule: ScalingRule!
+  replicaState: ReplicaState!
   revisionHistory(filter: ModelRevisionFilter = null, orderBy: [ModelRevisionOrderBy!] = null, before: String = null, after: String = null, first: Int = null, last: Int = null, limit: Int = null, offset: Int = null): ModelRevisionConnection!
 }
 
@@ -3046,10 +3046,10 @@ type ModelDeploymentMetadata
   name: String!
   status: DeploymentStatus!
   tags: [String!]!
-  project: GroupNode!
-  domain: DomainNode!
   createdAt: DateTime!
   updatedAt: DateTime!
+  project: GroupNode!
+  domain: DomainNode!
 }
 
 """Added in 25.13.0"""
@@ -3084,9 +3084,9 @@ input ModelDeploymentNetworkAccessInput
 type ModelMountConfig
   @join__type(graph: STRAWBERRY)
 {
-  vfolder: VirtualFolderNode!
   mountDestination: String!
   definitionPath: String!
+  vfolder: VirtualFolderNode!
 }
 
 """Added in 25.13.0"""
@@ -3105,7 +3105,6 @@ type ModelReplica implements Node
 {
   """The Globally Unique ID of this object"""
   id: ID!
-  revision: ModelRevision!
 
   """
   This represents whether the replica has been checked and its health state.
@@ -3138,6 +3137,7 @@ type ModelReplica implements Node
   The session ID associated with the replica. This can be null right after replica creation.
   """
   session: ComputeSessionNode!
+  revision: ModelRevision!
 }
 
 """Added in 25.13.0"""
@@ -3176,8 +3176,8 @@ type ModelRevision implements Node
   modelRuntimeConfig: ModelRuntimeConfig!
   modelMountConfig: ModelMountConfig!
   extraMounts: ExtraVFolderMountConnection!
-  image: ImageNode!
   createdAt: DateTime!
+  image: ImageNode!
 }
 
 """Added in 25.13.0"""
@@ -4103,7 +4103,9 @@ type Mutation
   """Added in 25.13.0"""
   addModelRevision(input: AddModelRevisionInput!): AddModelRevisionPayload! @join__field(graph: STRAWBERRY)
 
-  """Added in 25.13.0"""
+  """
+  Added in 25.13.0. Create model revision which is not attached to any deployment.
+  """
   createModelRevision(input: CreateModelRevisionInput!): CreateModelRevisionPayload! @join__field(graph: STRAWBERRY)
 
   """Added in 25.14.0"""
@@ -5066,8 +5068,6 @@ type ReservoirRegistryEdge
 type ResourceConfig
   @join__type(graph: STRAWBERRY)
 {
-  resourceGroup: ScalingGroupNode!
-
   """
   Resource Slots are a JSON string that describes the resources allocated for the deployment. Example: "resourceSlots": "{\"cpu\": \"1\", \"mem\": \"1073741824\", \"cuda.device\": \"0\"}"
   """
@@ -5077,6 +5077,7 @@ type ResourceConfig
   Resource Options are a JSON string that describes additional options for the resources. This is especially used for shared memory configurations. Example: "resourceOpts": "{\"shmem\": \"64m\"}"
   """
   resourceOpts: JSONString
+  resourceGroup: ScalingGroupNode!
 }
 
 """Added in 25.13.0"""

docs/manager/graphql-reference/v2-schema.graphql

@@ -710,10 +710,10 @@ type ModelDeployment implements Node {
   metadata: ModelDeploymentMetadata!
   networkAccess: ModelDeploymentNetworkAccess!
   revision: ModelRevision
-  scalingRule: ScalingRule!
-  replicaState: ReplicaState!
   defaultDeploymentStrategy: DeploymentStrategy!
   createdUser: UserNode!
+  scalingRule: ScalingRule!
+  replicaState: ReplicaState!
   revisionHistory(filter: ModelRevisionFilter = null, orderBy: [ModelRevisionOrderBy!] = null, before: String = null, after: String = null, first: Int = null, last: Int = null, limit: Int = null, offset: Int = null): ModelRevisionConnection!
 }
 
@@ -741,10 +741,10 @@ type ModelDeploymentMetadata {
   name: String!
   status: DeploymentStatus!
   tags: [String!]!
-  project: GroupNode!
-  domain: DomainNode!
   createdAt: DateTime!
   updatedAt: DateTime!
+  project: GroupNode!
+  domain: DomainNode!
 }
 
 """Added in 25.13.0"""
@@ -771,9 +771,9 @@ input ModelDeploymentNetworkAccessInput {
 
 """Added in 25.13.0"""
 type ModelMountConfig {
-  vfolder: VirtualFolderNode!
   mountDestination: String!
   definitionPath: String!
+  vfolder: VirtualFolderNode!
 }
 
 """Added in 25.13.0"""
@@ -787,7 +787,6 @@ input ModelMountConfigInput {
 type ModelReplica implements Node {
   """The Globally Unique ID of this object"""
   id: ID!
-  revision: ModelRevision!
 
   """
   This represents whether the replica has been checked and its health state.
@@ -820,6 +819,7 @@ type ModelReplica implements Node {
   The session ID associated with the replica. This can be null right after replica creation.
   """
   session: ComputeSessionNode!
+  revision: ModelRevision!
 }
 
 """Added in 25.13.0"""
@@ -851,8 +851,8 @@ type ModelRevision implements Node {
   modelRuntimeConfig: ModelRuntimeConfig!
   modelMountConfig: ModelMountConfig!
   extraMounts: ExtraVFolderMountConnection!
-  image: ImageNode!
   createdAt: DateTime!
+  image: ImageNode!
 }
 
 """Added in 25.13.0"""
@@ -961,7 +961,9 @@ type Mutation {
   """Added in 25.13.0"""
   addModelRevision(input: AddModelRevisionInput!): AddModelRevisionPayload!
 
-  """Added in 25.13.0"""
+  """
+  Added in 25.13.0. Create model revision which is not attached to any deployment.
+  """
   createModelRevision(input: CreateModelRevisionInput!): CreateModelRevisionPayload!
 
   """Added in 25.14.0"""
@@ -1272,8 +1274,6 @@ type ReservoirRegistryEdge {
 
 """Added in 25.13.0"""
 type ResourceConfig {
-  resourceGroup: ScalingGroupNode!
-
   """
   Resource Slots are a JSON string that describes the resources allocated for the deployment. Example: "resourceSlots": "{\"cpu\": \"1\", \"mem\": \"1073741824\", \"cuda.device\": \"0\"}"
   """
@@ -1283,6 +1283,7 @@ type ResourceConfig {
   Resource Options are a JSON string that describes additional options for the resources. This is especially used for shared memory configurations. Example: "resourceOpts": "{\"shmem\": \"64m\"}"
   """
   resourceOpts: JSONString
+  resourceGroup: ScalingGroupNode!
 }
 
 """Added in 25.13.0"""

src/ai/backend/common/exception.py

@@ -165,6 +165,7 @@ class ErrorDomain(enum.StrEnum):
     PERMISSION = "permission"
     METRIC = "metric"
     STORAGE_PROXY = "storage-proxy"
+    MODEL_DEPLOYMENT = "model-deployment"
 
 
 class ErrorOperation(enum.StrEnum):
@@ -643,3 +644,16 @@ def error_code(cls) -> ErrorCode:
             operation=ErrorOperation.READ,
             error_detail=ErrorDetail.NOT_FOUND,
         )
+
+
+class ModelDeploymentUnavailableError(BackendAIError, web.HTTPServiceUnavailable):
+    error_type = "https://api.backend.ai/probs/model-deployment-unavailable"
+    error_title = "Model Deployment Unavailable"
+
+    @classmethod
+    def error_code(cls) -> ErrorCode:
+        return ErrorCode(
+            domain=ErrorDomain.MODEL_DEPLOYMENT,
+            operation=ErrorOperation.EXECUTE,
+            error_detail=ErrorDetail.UNAVAILABLE,
+        )

src/ai/backend/manager/api/gql/base.py

@@ -5,6 +5,7 @@
 from enum import StrEnum
 from typing import TYPE_CHECKING, Any, Optional, Type, cast
 
+import graphene
 import orjson
 import strawberry
 from graphql import StringValueNode
@@ -160,7 +161,17 @@ def from_resource_slot(resource_slot: ResourceSlot) -> JSONString:
         return JSONString.serialize(resource_slot.to_json())
 
 
-def to_global_id(type_: Type[Any], local_id: uuid.UUID | str) -> str:
+def to_global_id(
+    type_: Type[Any], local_id: uuid.UUID | str, is_target_graphene_object: bool = False
+) -> str:
+    if is_target_graphene_object:
+        # For compatibility with existing Graphene-based global IDs
+        if not issubclass(type_, graphene.ObjectType):
+            raise TypeError(
+                "type_ must be a graphene ObjectType when is_target_graphene_object is True."
+            )
+        typename = type_.__name__
+        return base64(f"{typename}:{local_id}")
     if not has_object_definition(type_):
         raise TypeError("type_ must be a Strawberry object type (Node or Edge).")
     typename = get_object_definition(type_, strict=True).name

src/ai/backend/manager/api/gql/model_deployment/access_token.py

@@ -1,16 +1,27 @@
-from datetime import datetime, timedelta
+from collections.abc import Sequence
+from datetime import datetime
+from typing import Self
 from uuid import UUID
 
 import strawberry
 from strawberry import ID, Info
 from strawberry.relay import Connection, Edge, Node, NodeID
 
+from ai.backend.common.exception import ModelDeploymentUnavailableError
 from ai.backend.manager.api.gql.types import StrawberryGQLContext
+from ai.backend.manager.data.deployment.access_token import ModelDeploymentAccessTokenCreator
+from ai.backend.manager.data.deployment.types import ModelDeploymentAccessTokenData
+from ai.backend.manager.services.deployment.actions.access_token.create_access_token import (
+    CreateAccessTokenAction,
+)
+from ai.backend.manager.services.deployment.actions.access_token.get_access_tokens_by_deployment_id import (
+    GetAccessTokensByDeploymentIdAction,
+)
 
 
 @strawberry.type
 class AccessToken(Node):
-    id: NodeID
+    id: NodeID[str]
     token: str = strawberry.field(description="Added in 25.13.0: The access token.")
     created_at: datetime = strawberry.field(
         description="Added in 25.13.0: The creation timestamp of the access token."
@@ -19,6 +30,34 @@ class AccessToken(Node):
         description="Added in 25.13.0: The expiration timestamp of the access token."
     )
 
+    @classmethod
+    def from_dataclass(cls, data: ModelDeploymentAccessTokenData) -> Self:
+        return cls(
+            id=ID(str(data.id)),
+            token=data.token,
+            created_at=data.created_at,
+            valid_until=data.valid_until,
+        )
+
+    @classmethod
+    async def batch_load_by_deployment_ids(
+        cls, ctx: StrawberryGQLContext, deployment_ids: Sequence[UUID]
+    ) -> list[list[ModelDeploymentAccessTokenData]]:
+        """Batch load access tokens by deployment IDs."""
+        processor = ctx.processors.deployment
+        if processor is None:
+            raise ModelDeploymentUnavailableError(
+                "Model Deployment feature is unavailable. Please contact support."
+            )
+
+        results = []
+        for deployment_id in deployment_ids:
+            result = await processor.get_access_tokens_by_deployment_id.wait_for_complete(
+                GetAccessTokensByDeploymentIdAction(deployment_id=deployment_id)
+            )
+            results.append(result.data if result else [])
+        return results
+
 
 AccessTokenEdge = Edge[AccessToken]
 
@@ -32,42 +71,6 @@ def __init__(self, *args, count: int, **kwargs):
         self.count = count
 
 
-mock_access_token_1 = AccessToken(
-    id=UUID("13cd8325-9307-49e4-94eb-ded2581363f8"),
-    token="mock-token-1",
-    created_at=datetime.now(),
-    valid_until=datetime.now() + timedelta(hours=12),
-)
-
-mock_access_token_2 = AccessToken(
-    id=UUID("dc1a223a-7437-4e6f-aedf-23417d0486dd"),
-    token="mock-token-2",
-    created_at=datetime.now(),
-    valid_until=datetime.now() + timedelta(hours=1),
-)
-
-mock_access_token_3 = AccessToken(
-    id=UUID("39f8b49e-0ddf-4dfb-92d6-003c771684b7"),
-    token="mock-token-3",
-    created_at=datetime.now(),
-    valid_until=datetime.now() + timedelta(hours=100),
-)
-
-mock_access_token_4 = AccessToken(
-    id=UUID("85a6ed1e-133b-4f58-9c06-f667337c6111"),
-    token="mock-token-4",
-    created_at=datetime.now(),
-    valid_until=datetime.now() + timedelta(hours=10),
-)
-
-mock_access_token_5 = AccessToken(
-    id=UUID("c42f8578-b31d-4203-b858-93f93b4b9549"),
-    token="mock-token-5",
-    created_at=datetime.now(),
-    valid_until=datetime.now() + timedelta(hours=3),
-)
-
-
 @strawberry.input
 class CreateAccessTokenInput:
     model_deployment_id: ID = strawberry.field(
@@ -77,6 +80,12 @@ class CreateAccessTokenInput:
         description="Added in 25.13.0: The expiration timestamp of the access token."
     )
 
+    def to_creator(self) -> "ModelDeploymentAccessTokenCreator":
+        return ModelDeploymentAccessTokenCreator(
+            model_deployment_id=UUID(self.model_deployment_id),
+            valid_until=self.valid_until,
+        )
+
 
 @strawberry.type
 class CreateAccessTokenPayload:
@@ -87,4 +96,9 @@ class CreateAccessTokenPayload:
 async def create_access_token(
     input: CreateAccessTokenInput, info: Info[StrawberryGQLContext]
 ) -> CreateAccessTokenPayload:
-    return CreateAccessTokenPayload(access_token=mock_access_token_1)
+    deployment_processor = info.context.processors.deployment
+    assert deployment_processor is not None
+    result = await deployment_processor.create_access_token.wait_for_complete(
+        action=CreateAccessTokenAction(input.to_creator())
+    )
+    return CreateAccessTokenPayload(access_token=AccessToken.from_dataclass(result.data))